aquakin.optimize_design#

aquakin.optimize_design(objective, bounds, *, input_names=None, constraints=(), x0=None, maximize=False, method='SLSQP', n_starts=1, seed=0, tol=1e-06, constraint_tol=0.0001)[source]#

Minimise (or maximise) an objective over a bounded design space subject to inequality constraints, using autodiff gradients.

The canonical use is “size a design to a permit at minimum cost”: objective is an operational-cost / energy metric and each Constraint is an effluent-quality ceiling. objective and the constraint functions share the fn(x) -> scalar contract of monte_carlo() – they build the params / initial state and run the solve themselves – and must be JAX-differentiable, since their gradients are taken by autodiff and passed to the optimizer (a gradient-based, constrained NLP solver via SciPy).

Parameters:

  • objective (callable) – objective(x) -> scalar to minimise (or maximise; see maximize).

  • bounds (sequence of (low, high)) – Box bounds for each design variable, length d.

  • input_names (sequence of str, optional) – Design-variable names (defaults to x0..).

  • constraints (sequence of Constraint) – Inequality constraints lower <= c.fn(x) <= upper.

  • x0 (sequence of float, optional) – Starting point. Defaults to the box centre; with n_starts > 1 it is ignored in favour of quasi-random starts.

  • maximize (bool) – Maximise instead of minimise.

  • method (str) – SciPy constrained method (default "SLSQP"; "trust-constr" also works with bounds + constraints).

  • n_starts (int) – Multistart count – quasi-random (Sobol) starts in the box; the best feasible optimum is returned. Escapes local minima on multimodal designs.

  • seed (int) – Seed for the multistart sampler (reproducible).

  • tol (float) – Optimizer tolerance and the slack within which a constraint counts as satisfied when judging feasibility / picking the multistart winner.

  • constraint_tol (float) – Optimizer tolerance and the slack within which a constraint counts as satisfied when judging feasibility / picking the multistart winner.

Return type:

OptimizeResult